Microsoft Research releases a public preview of BioEmu2, an MIT-licensed open-weights foundation model for biomolecular structure and dynamics covering proteins, nucleic acids and ligands

Why it matters

【Key points】

On Oct 20 (09:13 UTC), Microsoft Research AI for Science released a public preview of BioEmu2, a foundation model for biomolecular structure and dynamics, with code and weights under MIT. Team lead Frank Noé announced it on X; checkpoint bioemu2-preview is on GitHub and Hugging Face. Its predecessor BioEmu (Science, 2025) handled single proteins only.

1. Scope: proteins, nucleic acids, ligands and other components; structure prediction and equilibrium sampling; megadalton complexes of 10,000+ residues on multiple GPUs; experimental data can be used as constraints.

2. Dynamics: implicit transfer operators move structures forward by nanoseconds, microseconds or to equilibrium. Noé calls it the only state-of-the-art biomolecular foundation model that predicts dynamics.

3. Team-reported results: structure prediction competitive with leading folding models on the same train/test split (Runs N' Poses, FoldBench); free energies within about 1 kcal/mol, covering 50-100% of MD structures; a 5 ms peptide-binding run that takes 10+ GPU-years with MD takes about 40 minutes, a 100,000x+ speedup.

4. Needs Linux and an NVIDIA GPU. No paper or official blog post yet.

【Verifiable】

GitHub: https://github.com/microsoft/bioemu2

Hugging Face: https://huggingface.co/microsoft/bioemu2

Announcement thread: https://x.com/FrankNoeBerlin/status/2112472173336240499

【Our read】

1. Adding nucleic acids, small molecules and dynamics targets how molecules move and bind, which static structure models do not cover; that matters for drug design.

2. All results are self-reported; the dynamics claim awaits peer review.

3. Open MIT weights make it easy to try; as a preview, it may change.